GPT-5.3-Codex Tops Coding Benchmarks, Minimax M2.5 and GLM 5 Challenge Open Weight Leaders

GPT-5.3-Codex Tops Coding Benchmarks, Minimax M2.5 and GLM 5 Challenge Open Weight Leaders

SanityBoard, the independent coding evaluation platform developed by researcher lemon07r, has released a major update showcasing unprecedented performance gains across leading AI coding models. The latest results, published on February 6, 2026, confirm that OpenAI’s newly released GPT-5.3-Codex has dethroned June CLI as the top-performing agentic coding system, leveraging its advanced subagent architecture to achieve record-breaking scores on complex programming tasks. Simultaneously, open-weight models Minimax M2.5 and GLM 5 have emerged as formidable contenders, challenging the dominance of previously top-ranked systems like Kimi K2.5.

According to TechSpot, GPT-5.3-Codex’s architectural breakthrough lies in its self-referential training methodology — the model reportedly helped refine its own code generation pipeline during development, enabling deeper reasoning over multi-step programming workflows. This self-improvement capability, combined with enhanced tool-use integration and context retention, has positioned GPT-5.3-Codex as the most capable coding agent to date. Fast Company adds that the model demonstrates a unique ability to think "deeper and wider," meaning it can simultaneously analyze multiple implementation strategies, anticipate edge cases, and dynamically switch between high-level architecture planning and low-level syntax optimization — a level of cognitive flexibility previously unseen in commercial coding models.

On the open-weight front, the Minimax M2.5 paired with the Droid agent has overtaken the Kimi K2.5 + Kimi CLI combo as the highest-performing open-source configuration. This shift underscores the growing importance of agent orchestration: while model size and training data remain critical, the choice of agent framework — the system that manages planning, tool selection, and error recovery — can dramatically amplify performance. Droid, known for its modular, state-aware task decomposition, appears to extract maximum potential from M2.5’s reasoning capabilities, enabling it to solve intricate algorithmic challenges with fewer iterations and higher accuracy than previous combinations.

Perhaps the most surprising development is GLM 5’s performance on the Opencode benchmark, where it achieved the highest score ever recorded among open-weight models. Although testing on Droid remains pending due to infrastructure limitations at ZAI Labs — the provider of the API endpoint — early results suggest GLM 5 may soon surpass both Minimax M2.5 and Kimi K2.5 in overall capability. The model’s strength appears rooted in its dense, multilingual code corpus and improved instruction-following architecture, making it particularly adept at interpreting ambiguous or incomplete problem specifications — a common hurdle in real-world software engineering.

Notably, newer iterations of Claude Code have improved Kimi K2.5’s performance, but showed minimal gains for Anthropic’s Opus 4.5, indicating that model-agent compatibility is highly non-linear. This reinforces the notion that there is no universal "best" model — optimal performance depends on the synergy between architecture, agent framework, and evaluation criteria.

The lead researcher behind SanityBoard has publicly cited API rate-limiting issues with ZAI Labs as the primary bottleneck preventing comprehensive testing of GLM 5 across multiple agent frameworks. With evaluations currently limited to Opencode due to frequent 5- to 15-minute delays between tasks, future benchmarks may reveal even more dramatic shifts in the leaderboard. Plans are underway to compare OpenAI’s endpoint against Anthropic’s using the same evaluation harness, which could provide the first apples-to-apples comparison of proprietary versus open-weight systems under identical conditions.

As the AI coding landscape evolves from isolated model evaluations to holistic agent-system assessments, SanityBoard has become an indispensable resource for developers, researchers, and AI engineers. With open-source repositories for both the evaluation harness and leaderboard publicly available on GitHub, the community is invited to replicate, extend, and challenge these findings — a rare and commendable commitment to transparency in an increasingly proprietary AI ecosystem.